基于深度半监督的柬语句子级情感分类

针对柬语标注数据较少、语料稀缺,柬语句子级情感分析任务进步缓慢的问题,提出了一种基于深度半监督CNN(convolutional neural networks)的柬语句子级情感极性分类方法.该方法通过融合词典嵌入的分开卷积CNN模型,利用少量已有的柬语情感词典资源提升句子级情感分类任务性能.首先构建柬语句子词嵌入和词典嵌入,通过使用不同的卷积核对两部分嵌入分别进行卷积,将已有情感词典信息融入到CNN模型中去,经过最大延时池化得到最大输出特征,把两部分最大输出特征拼接后作为全连接层输入;然后通过结合半监督学习方法——时序组合模型,训练提出的深度神经网络模型,利用标注与未标注语料训练,降低对标注语料的需求,进一步提升模型情感分类的准确性.结果 证明,通过半监督方法时序组合模型训练,在人工标记数据相同的情况下,该方法相较于监督方法在柬语句子级情感分类任务上准确率提升了3.89％.

Sentiment classification of Khmer sentences based on deep semi-supervised

Aiming at the problems of limited annotation data, scarce corpus and slow progress of Khmer sentence-level sentiment analysis, this paper proposed a Khmer sentence-level sentiment classification method based on deep semi-supervised CNN model. This method combined the separate convolution for word and lexicon embeddings, used a small amount of existing Khmer sentiment lexicon resources to improve sentence-level sentiment classification task performance. First, it constructed the word and lexicon embeddings of Khmer sentence, used different convolution kernels to convolve two-part embeddings respectively, integrated the existing sentiment lexicon information into the CNN model. After the max-over-time pooling, obtaining the maximum output feature. The maximum output features of the two parts were stitched together as the input of the full connection layer. And then, it used the semi-supervised learning method of temporal ensembling training the deep neural network, reduced the need for annotated corpus, and further improved the accuracy of the model''s sentiment classification. The result proves that through the semi-supervised method of temporal ensembling model training, the accuracy of this method is 3.89% higher than that of the supervised method in the Khmer sentence-level sentiment classification task when the artificially labeled data is the same.